Early Academic Risk Prediction for Personalized Learning Support Using Course-Normalized Assessment-Aware Temporal Fusion

Early academic risk prediction can support timely advising by identifying learners likely to fail or withdraw before course completion. We propose CATEF, a course-normalized assessment-aware temporal fusion framework that combines CATEF-Net with a LightGBM branch through validation-selected probability fusion. CATEF-Net models weekly activity together with learner and assessment context, while LightGBM captures nonlinear interactions in flattened structured features. Experiments on the Open University Learning Analytics Dataset use prediction cutoffs at weeks 4, 8, 12, and 16, grouped student-level splits, five random seeds, and validation-only selection of fusion weights and decision thresholds. Across 20 test runs per method, CATEF achieves a mean Risk-F1 of 0.8235, a precision–recall area under the curve (PR-AUC) of 0.9235, and a receiver operating characteristic area under the curve (ROC-AUC) of 0.8990. Risk-F1 gains over the strongest baselines are modest, whereas ranking improvements are more consistent. Ablation identifies assessment information as the dominant predictive signal. CATEF is therefore best viewed as an early-risk screening and support-prioritization layer rather than an automatic intervention system.